Mean reversion rsi

RSI (Relative Strength Index)

Bullish: RSI crosses above oversold level (default 30) from below. Bearish: RSI crosses below overbought level (default 70) from above. Uses Wilder's smoothed moving average.

Signal family

Mean reversion — Oscillator-based signals that fire at overbought or oversold extremes — typically fade the prevailing move.

Parameters

Name Description Default Range
period RSI period 14 5–50
overbought Overbought level 70 60–90
oversold Oversold level 30 10–40

Historical context

1,194,048 triggers on 23,777 tickers, 1988-04-29 → 2026-05-01. Universe: global — all covered exchanges (mcap ≥ $100,000,000, price ≥ $1). Long-only convention: BUY at open T+1, hold the horizon, compare to S&P 500 Equal Weight over the same window.

Methodology footnotes

Benchmarks shown in the detail tables: spxew (S&P 500 Equal Weight — primary, median-stock view, avoids the 2020+ megacap-concentration distortion), spx (S&P 500 cap-weighted, distorted post-2020), msci (MSCI World USD). Per-stock regime tags: trending = ADX(14) ≥ 25, high vol = 20d realized annualized vol ≥ 20%. 1d return = intraday T+1 open→close; 20d = open T+1 to close T+20.

At a glance — alpha vs S&P 500 Equal Weight, global universe

Holding-period sensitivity. Bullish columns: positive = signal worked (long the trigger beat the index). Bearish columns: negative = signal worked (the flagged stock underperformed).

Horizon Bullish α Bearish α
5-day +0.19% +0.07%
20-day +0.25% +0.47%
60-day +0.22% +1.05%
1-year +1.14% +5.79%
Random-date null check (20-day): Bullish: beats random (p=0.005).
Bearish: worse than random (p=1.000).

Where does RSI actually fire?

The bucket distribution often reveals what the signal really is, regardless of its textbook label. Heavy concentration in "non-trending + high vol" = it's mostly a chop-market event. Heavy in "trending + low vol" = it picks up the smooth grinds. Read the chart before the alpha numbers — context shapes everything that follows.

RSI (Relative Strength Index) (rsi) — trigger count distribution by per-stock regime quadrant (trending/non-trending × high/low realized volatility) for bullish and bearish triggers, global universe

Does it work in every regime?

Trigger alpha split by the host stock's own regime on the trigger date — trending or ranging, high-vol or low-vol. The 20d alpha you'd actually capture if you took the trade. Bars matching your direction's "right" sign (positive for bullish, negative for bearish) = the signal worked in that regime; opposite sign = avoid it there. A signal with one strong-positive bar and three flat ones isn't a "20d alpha" signal — it's a "20d alpha when the stock is X" signal.

RSI (Relative Strength Index) (rsi) — mean 20-day alpha versus S&P 500 Equal Weight by per-stock regime quadrant, bullish and bearish triggers side by side
Trending + Low vol
Stock in a clean directional move with low realized volatility. Textbook "trend-following paradise" — smooth grind with little whipsaw risk.
Trending + High vol
Violent directional moves — parabolic rallies, crisis selloffs. Trend exists but the path is noisy. Signal timing may be imprecise.
Non-trending + Low vol
Quiet chop, summer doldrums, consolidations. No directional bias but also no big swings — small edges become reliable if they exist at all.
Non-trending + High vol
Choppy and violent — the classical "whipsaw zone" for momentum signals. Crossovers and breakouts fire repeatedly without follow-through.

Does it work in every era?

A multi-year average can hide major instability. The sample splits into three windows: 2015–2019 (pre-COVID), 2020–2022 (pandemic + 2022 bear), and 2023+ (post-ZIRP + AI megacap rally). All three matching your direction's "right" sign = the signal is durable. One era doing all the work = a regime-specific edge that may not repeat. The bigger the variance across eras, the smaller the position you should run.

RSI (Relative Strength Index) (rsi) — 20-day alpha split by historical sub-period (2015-2019, 2020-2022, 2023+) to check consistency across market regimes

↑ Bullish triggers

Bench Metric 1d 5d 20d 60d 252d
spxew Stock % +0.09% +0.47% +1.78% +3.66% +13.30%
Bench % +0.06% +0.21% +1.47% +3.44% +11.68%
Alpha % +0.01% +0.19% +0.25% +0.22% +1.14%
Median alpha -0.07% -0.08% -0.54% -1.78% -7.38%
Hit rate (α>0) 48.3% 49.2% 47.4% 45.1% 40.9%
p (naive) 0.0722 <0.001 <0.001 <0.001 <0.001
p (HAC) 0.0731 <0.001 <0.001 <0.001 <0.001
N 495,565 479,214 472,184 460,098 427,228
spx Stock % +0.09% +0.47% +1.78% +3.66% +13.30%
Bench % +0.06% +0.29% +1.74% +3.84% +15.48%
Alpha % +0.00% +0.13% +0.03% -0.18% -2.13%
Median alpha -0.06% -0.13% -0.76% -2.19% -11.03%
Hit rate (α>0) 48.6% 48.7% 46.3% 44.2% 37.4%
p (naive) 0.8919 <0.001 0.1374 <0.001 <0.001
p (HAC) 0.8920 <0.001 0.2332 <0.001 <0.001
N 497,854 482,511 481,046 465,046 436,805
msci Stock % +0.09% +0.47% +1.78% +3.66% +13.30%
Bench % +0.09% +0.29% +1.50% +3.54% +13.03%
Alpha % +0.01% +0.13% +0.20% +0.15% -0.54%
Median alpha -0.09% -0.16% -0.60% -1.92% -9.02%
Hit rate (α>0) 48.0% 48.4% 47.0% 44.8% 39.2%
p (naive) 0.1346 <0.001 <0.001 <0.001 <0.001
p (HAC) 0.1361 <0.001 <0.001 0.0005 0.0087
N 494,500 477,152 471,477 462,324 426,023
Observed 20-day lift (vertical line) against the null distribution of random-date firing. If the line is deep inside the null cloud, the signal adds no information. If it sits in the right tail, the signal is doing real work in that direction; in the left tail it ran inverted — random dates served the bullish case better than its own triggers did.
RSI (Relative Strength Index) (rsi) — bullish 20-day observed lift versus random-date permutation null (200 iterations)
Permutation null detail — all horizons × each benchmark
200-iteration null: for each ticker, sample N random dates from its history (matching observed trigger count) and compute the same alpha. Both observed and null are baseline-centered per ticker (each ticker's own baseline alpha is subtracted), which removes the universe-selection lift that all surviving names share, so the comparison is about the trigger's timing. It does not put the null at zero: that baseline is a median, while a random date's expectation is the ticker's mean, so the null settles at the gap between the two — the right-skew of the ticker's own alpha distribution, positive for equities. Read observed minus null, not the absolute position of either column — which is also why these figures do not match the α columns in the tables above (those are raw, uncentered means). pperm = one-sided fraction of null iters with mean in the "signal was right" tail (right for bullish, left for bearish), so it runs from a floor of 0.005 (trigger dates beat every random draw in the claimed direction) to a 1.000 ceiling (every random draw served the claimed direction better — a reliably inverted signal, not an absent one). The floor is 1/(iterations + 1) and moves with the iteration count; the ceiling does not. Either end says the result was reliable, not that it was large: the gap between the line and the cloud can be a fraction of a percent and still reach an endpoint, and the transaction-cost floor in the caveats below would swallow a gap that small. Read the size off the α and hit-rate columns, the reliability off pperm.
Horizon Bench Observed lift Null mean Null 95% CI pperm
1d spxew +0.16% +0.08% [+0.07%, +0.09%] 0.005
1d spx +0.14% +0.09% [+0.08%, +0.10%] 0.005
1d msci +0.17% +0.09% [+0.09%, +0.10%] 0.005
5d spxew +0.61% +0.36% [+0.34%, +0.38%] 0.005
5d spx +0.60% +0.38% [+0.36%, +0.39%] 0.005
5d msci +0.59% +0.38% [+0.36%, +0.40%] 0.005
20d spxew +1.48% +1.17% [+1.13%, +1.21%] 0.005
20d spx +1.49% +1.20% [+1.17%, +1.23%] 0.005
20d msci +1.54% +1.21% [+1.18%, +1.25%] 0.005
60d spxew +2.92% +2.54% [+2.49%, +2.60%] 0.005
60d spx +3.26% +2.61% [+2.55%, +2.66%] 0.005
60d msci +3.18% +2.63% [+2.57%, +2.69%] 0.005
252d spxew +6.00% +5.10% [+5.00%, +5.23%] 0.005
252d spx +6.63% +5.42% [+5.31%, +5.55%] 0.005
252d msci +6.15% +5.37% [+5.26%, +5.50%] 0.005

Example triggers on US large-caps (2023+, mcap ≥ $30B)

Six recent bullish RSI triggers on US mega-caps. Top three: the signal's best outcomes. Bottom three: the worst. Extreme outliers (|α| > 25%) excluded. The three best and three worst are still tail outcomes by construction — read them as the range, not the typical result.

Strongest outcomes (what RSI looks like when it works)
RSI (Relative Strength Index) (rsi) — strongest bullish trigger #1 on a US mega-cap, showing the stock's OHLC candles, moving averages, signal-specific overlays, and the 20-day forward window RSI (Relative Strength Index) (rsi) — strongest bullish trigger #2 on a US mega-cap, showing the stock's OHLC candles, moving averages, signal-specific overlays, and the 20-day forward window RSI (Relative Strength Index) (rsi) — strongest bullish trigger #3 on a US mega-cap, showing the stock's OHLC candles, moving averages, signal-specific overlays, and the 20-day forward window
Weakest outcomes (what RSI looks like when it fails)
RSI (Relative Strength Index) (rsi) — weakest bullish trigger #1 on a US mega-cap, showing the 20-day forward window where the signal failed RSI (Relative Strength Index) (rsi) — weakest bullish trigger #2 on a US mega-cap, showing the 20-day forward window where the signal failed RSI (Relative Strength Index) (rsi) — weakest bullish trigger #3 on a US mega-cap, showing the 20-day forward window where the signal failed
Stock-regime quadrants (2×2 per-stock, 20d alpha detail table)
Each quadrant groups triggers by the stock's own ADX(14) and RV(20) at the trigger date — the textbook conditioning variable (not market-level). Stock %, bench %, alpha %, and HAC p-value shown for each benchmark.
Quadrant N Stock % (spxew) Bench % (spxew) Alpha % (spxew) p (HAC) Stock % (spx) Bench % (spx) Alpha % (spx) p (HAC) Stock % (msci) Bench % (msci) Alpha % (msci) p (HAC)
Trending + Low vol Clean directional grind, low whipsaw 55,247 +0.22% +0.96% -0.67% <0.001 +0.22% +1.18% -0.93% <0.001 +0.22% +1.03% -0.79% <0.001
Trending + High vol Crisis selloff or parabolic rally 331,737 +2.19% +1.62% +0.50% <0.001 +2.19% +1.88% +0.29% <0.001 +2.19% +1.62% +0.49% <0.001
Non-trending + Low vol Quiet chop, summer doldrums 22,625 +0.01% +0.97% -0.90% <0.001 +0.01% +1.18% -1.14% <0.001 +0.01% +1.06% -1.03% <0.001
Non-trending + High vol Classical "whipsaw zone" for momentum 102,998 +1.74% +1.46% +0.22% <0.001 +1.74% +1.79% -0.04% 0.3339 +1.74% +1.59% +0.08% 0.0715
Sub-period breakdown table (20d alpha)
Historical clustering check. If alpha concentrates in one era, the signal's robustness is questionable.
Period N Alpha % (spxew) p (HAC) Alpha % (spx) p (HAC) Alpha % (msci) p (HAC)
2015-2019 2015-01-01 → 2020-01-01 164,305 +0.13% <0.001 +0.16% <0.001 +0.39% <0.001
2020-2022 2020-01-01 → 2023-01-01 154,976 +0.19% <0.001 +0.18% <0.001 +0.48% <0.001
2023-2026 2023-01-01 → 2099-01-01 193,159 +0.41% <0.001 -0.21% <0.001 -0.21% <0.001

↓ Bearish triggers negative alpha = signal was right (stock underperformed market)

Bench Metric 1d 5d 20d 60d 252d
spxew Stock % +0.03% +0.24% +1.05% +3.09% +15.48%
Bench % +0.03% +0.17% +0.65% +2.07% +9.76%
Alpha % -0.01% +0.07% +0.47% +1.05% +5.79%
Median alpha -0.08% -0.27% -0.69% -1.71% -4.75%
Hit rate (α>0) 48.2% 47.3% 46.7% 45.5% 44.3%
p (naive) 0.1552 <0.001 <0.001 <0.001 <0.001
p (HAC) 0.1555 <0.001 <0.001 <0.001 <0.001
N 654,527 632,314 630,024 612,381 534,017
spx Stock % +0.03% +0.24% +1.05% +3.09% +15.48%
Bench % +0.01% +0.19% +0.89% +3.01% +13.67%
Alpha % +0.01% +0.05% +0.20% +0.08% +1.72%
Median alpha -0.07% -0.31% -1.00% -2.74% -8.91%
Hit rate (α>0) 48.2% 46.8% 45.2% 42.8% 39.8%
p (naive) 0.0008 <0.001 <0.001 0.0055 <0.001
p (HAC) 0.0008 <0.001 <0.001 0.1315 <0.001
N 660,384 640,062 635,556 620,955 539,837
msci Stock % +0.03% +0.24% +1.05% +3.09% +15.48%
Bench % +0.03% +0.17% +0.75% +2.55% +11.10%
Alpha % -0.00% +0.08% +0.35% +0.56% +4.25%
Median alpha -0.08% -0.29% -0.86% -2.27% -6.35%
Hit rate (α>0) 48.0% 47.0% 45.9% 44.0% 42.6%
p (naive) 0.7245 <0.001 <0.001 <0.001 <0.001
p (HAC) 0.7247 <0.001 <0.001 <0.001 <0.001
N 656,764 637,264 634,670 615,332 537,887
Observed 20-day lift (vertical line) against the null distribution of random-date firing. If the line is deep inside the null cloud, the signal adds no information. If it sits in the left tail, the signal is doing real work in that direction; in the right tail it ran inverted — random dates served the bearish case better than its own triggers did.
RSI (Relative Strength Index) (rsi) — bearish 20-day observed lift versus random-date permutation null (200 iterations)
Permutation null detail — all horizons × each benchmark
200-iteration null: for each ticker, sample N random dates from its history (matching observed trigger count) and compute the same alpha. Both observed and null are baseline-centered per ticker (each ticker's own baseline alpha is subtracted), which removes the universe-selection lift that all surviving names share, so the comparison is about the trigger's timing. It does not put the null at zero: that baseline is a median, while a random date's expectation is the ticker's mean, so the null settles at the gap between the two — the right-skew of the ticker's own alpha distribution, positive for equities. Read observed minus null, not the absolute position of either column — which is also why these figures do not match the α columns in the tables above (those are raw, uncentered means). pperm = one-sided fraction of null iters with mean in the "signal was right" tail (right for bullish, left for bearish), so it runs from a floor of 0.005 (trigger dates beat every random draw in the claimed direction) to a 1.000 ceiling (every random draw served the claimed direction better — a reliably inverted signal, not an absent one). The floor is 1/(iterations + 1) and moves with the iteration count; the ceiling does not. Either end says the result was reliable, not that it was large: the gap between the line and the cloud can be a fraction of a percent and still reach an endpoint, and the transaction-cost floor in the caveats below would swallow a gap that small. Read the size off the α and hit-rate columns, the reliability off pperm.
Horizon Bench Observed lift Null mean Null 95% CI pperm
1d spxew +0.13% +0.08% [+0.07%, +0.08%] 1.000
1d spx +0.13% +0.09% [+0.08%, +0.09%] 1.000
1d msci +0.15% +0.09% [+0.08%, +0.10%] 1.000
5d spxew +0.42% +0.35% [+0.33%, +0.36%] 1.000
5d spx +0.44% +0.36% [+0.35%, +0.38%] 1.000
5d msci +0.46% +0.37% [+0.36%, +0.39%] 1.000
20d spxew +1.34% +1.14% [+1.11%, +1.17%] 1.000
20d spx +1.31% +1.17% [+1.14%, +1.20%] 1.000
20d msci +1.34% +1.18% [+1.15%, +1.22%] 1.000
60d spxew +2.42% +2.48% [+2.43%, +2.52%] 0.030
60d spx +2.20% +2.54% [+2.49%, +2.59%] 0.005
60d msci +2.26% +2.56% [+2.51%, +2.60%] 0.005
252d spxew +4.48% +4.60% [+4.51%, +4.70%] 0.010
252d spx +4.58% +4.96% [+4.88%, +5.05%] 0.005
252d msci +4.72% +4.90% [+4.81%, +5.00%] 0.005

Example triggers on US large-caps (2023+, mcap ≥ $30B)

Six recent bearish RSI triggers on US mega-caps. Top three: the signal's best outcomes. Bottom three: the worst. Extreme outliers (|α| > 25%) excluded. The three best and three worst are still tail outcomes by construction — read them as the range, not the typical result.

Strongest outcomes (what RSI looks like when it works)
RSI (Relative Strength Index) (rsi) — strongest bearish trigger #1 on a US mega-cap, showing the stock's OHLC candles, moving averages, signal-specific overlays, and the 20-day forward window RSI (Relative Strength Index) (rsi) — strongest bearish trigger #2 on a US mega-cap, showing the stock's OHLC candles, moving averages, signal-specific overlays, and the 20-day forward window RSI (Relative Strength Index) (rsi) — strongest bearish trigger #3 on a US mega-cap, showing the stock's OHLC candles, moving averages, signal-specific overlays, and the 20-day forward window
Weakest outcomes (what RSI looks like when it fails)
RSI (Relative Strength Index) (rsi) — weakest bearish trigger #1 on a US mega-cap, showing the 20-day forward window where the signal failed RSI (Relative Strength Index) (rsi) — weakest bearish trigger #2 on a US mega-cap, showing the 20-day forward window where the signal failed RSI (Relative Strength Index) (rsi) — weakest bearish trigger #3 on a US mega-cap, showing the 20-day forward window where the signal failed
Stock-regime quadrants (2×2 per-stock, 20d alpha detail table)
Each quadrant groups triggers by the stock's own ADX(14) and RV(20) at the trigger date — the textbook conditioning variable (not market-level). Stock %, bench %, alpha %, and HAC p-value shown for each benchmark.
Quadrant N Stock % (spxew) Bench % (spxew) Alpha % (spxew) p (HAC) Stock % (spx) Bench % (spx) Alpha % (spx) p (HAC) Stock % (msci) Bench % (msci) Alpha % (msci) p (HAC)
Trending + Low vol Clean directional grind, low whipsaw 91,414 +0.43% +0.32% +0.15% <0.001 +0.43% +0.66% -0.22% <0.001 +0.43% +0.49% -0.04% 0.1218
Trending + High vol Crisis selloff or parabolic rally 473,601 +1.24% +0.70% +0.62% <0.001 +1.24% +0.94% +0.35% <0.001 +1.24% +0.80% +0.49% <0.001
Non-trending + Low vol Quiet chop, summer doldrums 28,011 +0.57% +0.38% +0.21% <0.001 +0.57% +0.66% -0.08% 0.0473 +0.57% +0.50% +0.09% 0.0401
Non-trending + High vol Classical "whipsaw zone" for momentum 88,401 +1.00% +0.77% +0.27% <0.001 +1.00% +0.94% +0.09% 0.0583 +1.00% +0.84% +0.20% <0.001
Sub-period breakdown table (20d alpha)
Historical clustering check. If alpha concentrates in one era, the signal's robustness is questionable.
Period N Alpha % (spxew) p (HAC) Alpha % (spx) p (HAC) Alpha % (msci) p (HAC)
2015-2019 2015-01-01 → 2020-01-01 205,195 -0.07% 0.0157 -0.37% <0.001 -0.15% <0.001
2020-2022 2020-01-01 → 2023-01-01 195,060 +0.31% <0.001 +0.48% <0.001 +0.60% <0.001
2023-2026 2023-01-01 → 2099-01-01 280,969 +1.01% <0.001 +0.45% <0.001 +0.56% <0.001

Methodology and caveats

How to read. Entry at open of T+1 (one trading day after the signal fires on close of T). 20d = open T+1 to close T+20. Alpha = stock return − benchmark return over the same window (Convention A, single-sided, textbook). For bullish triggers, POSITIVE alpha = signal was right. For bearish triggers, NEGATIVE alpha = signal was right (stock underperformed market). No sign-flipping; the direction of the bet determines what "good" looks like. Per-stock regime is each stock's own ADX(14) and RV(20) at the trigger date — not market-wide state.

Three p-values, three robustness tests. (a) p_naive: scipy one-sample t-test on winsorized alphas. Optimistic because overlapping 20d windows on the same ticker inflate effective N. (b) p_hac: Newey-West HAC with lag = horizon — corrects for the overlap and is the academic-finance standard. (c) p_perm: one-sided fraction of 200 random-date null iterations falling in the “signal was right” tail (mean ≥ observed for bullish; mean ≤ observed for bearish). Tests whether the signal beats random date selection at all. A signal that clears all three (pnaive, phac, pperm all < 0.05) has real information; a signal that fails pperm has not beaten random timing whatever the t-test says — and because the test is one-sided, a pperm up at its 1.000 ceiling is not "no edge" but inverted edge: every random draw served the claimed direction better than the trigger dates did.

Caveats. (i) Universe reflects today's active tickers; delisted losers pruned → survivorship bias. (ii) Mcap ≥ $100M filter uses today's snapshot, not point-in-time — mild lookahead on which stocks enter the sample, not on returns. (iii) Means and p-values use winsorized alphas (1/99 percentile) to prevent data errors from dominating. Medians and hit rates use raw data. (iv) Zero transaction costs assumed. Realistic bid-ask + commissions remove 20–40bps from 20d alpha on US large-caps, more on small-cap. Sub-20bps alpha is noise in practice. (v) Past performance does not predict future results.

How to use this

1 · When to reach for this signal

Use RSI (Relative Strength Index) bullish as a long-side screening tile. Bullish 20d alpha is +0.25% and beats random (permutation test, 200 iterations). Bearish 20d alpha is +0.47% — worse than random : firing on random dates would have done better.

These verdicts are 20-day holds vs S&P 500 Equal Weight. Longer horizons can differ in either direction — check the permutation detail tables below before extrapolating.

2 · When it works — the setups that drive it

  • Best bullish setup: Trending + High vol — alpha +0.50% / 20d on 331,737 historical triggers.
  • Least-bad bearish cell: Trending + Low vol — alpha +0.15% / 20d on 91,414 triggers — still wrong-signed; no bearish cell produced negative alpha.
  • Best era for bullish: 2023-2026 — alpha +0.41% / 20d on 193,159 triggers.
  • Best era for bearish: 2015-2019 — alpha -0.07% / 20d on 205,195 triggers.

3 · When it fails — common false positives

  • Weakest bullish cell: Non-trending + Low vol — alpha -0.90% / 20d on 22,625 triggers.
  • Weakest bearish cell: Trending + High vol — alpha +0.62% / 20d on 473,601 triggers.
  • Worst era for bullish: 2015-2019 — alpha +0.13% / 20d on 164,305 triggers.
  • Worst era for bearish: 2023-2026 — alpha +1.01% / 20d on 280,969 triggers.

Signal-specific failure patterns

Bullish fires on beaten-down names — the pool is adversely selected
The bullish trigger is RSI (14-period, Wilder's smoothing) crossing back above the 30 oversold line from below, which by construction selects stocks that just spent roughly two weeks dominated by down closes. In markets where returns concentrate in a narrow set of leaders, that pool is heavy with names underperforming for identifiable reasons — deteriorating fundamentals, sector de-ratings — and the anticipated oversold bounce often fails to complete. How this side is scoring right now is a question for the at-a-glance table and sub-period breakdown above, not for the mechanics; the mechanics only explain why the bullish side is the more regime-sensitive of the two directions.
Strong trends embed the oscillator — the exit cross is late by construction
In a persistent move RSI can pin inside the overbought or oversold zone for weeks (oscillator embedding). Both triggers fire only on the cross back out of the zone — bearish when RSI crosses back below 70 from above, bullish when it crosses back above 30 from below — so each fire arrives after the momentum extreme has already cooled. On the bearish side this cuts both ways: a healthy uptrend routinely dips under 70 during consolidation and then re-embeds, so many bearish fires mark pauses rather than tops. The regime-quadrant charts above show how outcomes currently differ between trending and non-trending host stocks.
Mean reversion is regime-dependent — read the sub-period rows before weighting either side
Every oscillator mean-reversion thesis assumes stretched prices snap back toward a stable anchor. Liquidity-driven markets break that assumption: when broad stimulus is re-flating everything, overbought names keep rising and oversold names keep falling because the anchor itself is moving. Sub-period results for signals in this family have flipped sign across macro eras, so a full-window average is a blend of dissimilar environments. Use the sub-period breakdown above as the current record; a verdict drawn from the pooled row alone will overstate stability.

4 · Pairing inside a screen

The statements below describe how this signal relates to others by construction — which indicator family it belongs to, and where same-family redundancy might reduce the independence of evidence inside a Daily Report. These are taxonomic classifications drawn from standard technical-analysis texts; they are not pairing backtests. Measured pair results — same-day co-fires put through the pair backtest — follow under “Measured pairings” below.

Oscillator-family redundancy

RSI belongs to the momentum-oscillator family alongside Stochastics, Williams %R, and CCI — each is a short-lookback price oscillator read against fixed reference bands (their constructions differ — RSI smooths close-to-close changes, Stochastics and Williams %R place the close inside the recent high-low range, CCI scales typical price around its own mean — but they respond to the same underlying move) (Murphy, Technical Analysis of the Financial Markets, 1999; Pring, Technical Analysis Explained, 5th ed. 2014; Kirkpatrick & Dahlquist, Technical Analysis, 3rd ed. 2015). Stacking two or more of these in the same direction within a single Daily Report produces correlated rather than independent evidence.

Measured pairings — Bonferroni survivors

Beyond the literature pairings above, these are the same-day co-fire combinations involving RSI (Relative Strength Index) that cleared the pair backtest's Bonferroni cut on the full 2016–2026 sample (549 pairs × 5 horizons = 2,745 hypotheses), on universes filtered to ADV ≥ $5M, price ≥ $5 and market cap ≥ $100M. That cut is two-sided: it asks only whether the co-fire's α is reliably different from zero, in either direction, so a pair can survive by reliably underperforming — 2 of the 27 rows below do exactly that (negative full-sample α). The same run holds out 2023+: the Test columns are that held-out window, printed for every row with enough held-out co-fires to measure, so a survivor that did not repeat out of sample is visible rather than hidden. All α figures here are for holding the stock long after the co-fire — no shorting assumed, and no sign flip for bearish legs. So positive α means the co-fire was followed by outperformance and negative α by underperformance, whichever way either leg points — a bearish leg does not flip the reading. Survivors are rare by design — absence of a pair here means it did not clear the cut, not that it was untested. Ranked by held-out (2023+) α. Historical tendencies, not recommendations.

US (NYSE / NASDAQ / AMEX)

Pair (same-day co-fire, long) Full α Test α (2023+) Test N p_perm test
rsi bullish + weekly_change bearish +1.10% +0.97% 1,931 0.004

1 of this universe's 18 surviving pairs involves this signal · α vs ^SPXEW.

Europe — 5 surviving pairs
Pair (same-day co-fire, long) Full α Test α (2023+) Test N p_perm test
macd bearish + rsi bearish +0.77% +0.80% 800 0.006
cci bearish + rsi bearish +0.60% +0.77% 3,073 0.002
rsi bearish + williams_r bearish +0.48% +0.52% 5,802 0.002
rsi bearish + stochastics bearish +0.41% +0.48% 3,779 0.002
hh_hl_structure bearish + rsi bullish +4.18% — — —

5 of this universe's 20 surviving pairs involve this signal · α vs ^STOXX.

Hong Kong — 4 surviving pairs
Pair (same-day co-fire, long) Full α Test α (2023+) Test N p_perm test
bollinger bearish + rsi bearish +1.41% +1.92% 1,280 0.002
rsi bearish + williams_r bearish +1.03% +1.91% 1,981 0.002
rsi bearish + stochastics bearish +1.19% +1.72% 1,276 0.002
rsi bearish + weekly_change bullish +2.14% +1.63% 498 0.024

4 of this universe's 23 surviving pairs involve this signal · α vs ^HSI.

China A-shares — 17 surviving pairs
Pair (same-day co-fire, long) Full α Test α (2023+) Test N p_perm test
rsi bullish + weekly_change bullish +4.70% +11.50% 207 0.002
hh_hl_streak bullish + rsi bullish +5.43% +10.50% 210 0.002
rsi bullish + weekly_change bearish +6.13% +6.96% 3,592 0.002
rsi bullish + volume_breakout bullish +1.45% +3.09% 2,122 0.002
cci bullish + rsi bullish +2.33% +3.00% 7,856 0.002
failed_double_top bearish + rsi bearish +2.12% +2.86% 749 0.002
bollinger bullish + rsi bullish +1.96% +1.38% 13,489 0.002
hh_hl_streak bearish + rsi bullish +1.09% +1.36% 889 0.002
rsi bearish + stochastics bearish +0.70% +1.21% 15,894 0.002
rsi bullish + williams_r bullish +1.85% +1.18% 19,261 0.002
new_20d_low bearish + rsi bullish +2.15% +1.08% 411 0.090
rsi bearish + williams_r bearish +0.43% +0.99% 27,909 0.002
cci bearish + rsi bearish +0.88% +0.92% 8,915 0.002
macd bullish + rsi bullish +1.09% +0.87% 1,588 0.002
rsi bullish + stochastics bullish +1.60% +0.81% 14,370 0.002
rsi bearish + volume_breakout bearish -0.53% +0.55% 11,970 0.002
new_20d_low bearish + rsi bearish -3.37% — — —

17 of this universe's 138 surviving pairs involve this signal · α vs 83188.HK.

China A-share survivor α runs large but skews toward small-caps, where trading costs and thin liquidity claim a large share of any measured edge — screening context, not a capturable spread.

“—” in the test columns means the held-out 2023+ sample fell below the 20-observation minimum this run requires before it computes any statistic, so no out-of-sample figure exists for that pair — not that it never co-fired again. Those pairs rank last.

What would likely rescue this signal

This block calls out the data or conditions that could turn a technically weak signal into a usable one in a composite screen. Based on signal mechanics and the observed failure patterns above; individual combinations are not yet backtested.

  • Regime-gate the bullish side — Bullish RSI selects recent losers, and mean reversion of losers has historically been environment-sensitive. Conditioning bullish fires on a market-level gate — breadth above a threshold, or the benchmark not sitting within a few percent of its high — is a testable way to isolate the environments where oversold bounces actually complete. Hypothesis, not yet run.
  • Fundamental filter for bullish mean-reversion — An oversold stock with an improving business is a different population from one that is falling because the business is deteriorating. The fundamentals filter in the report builder (live since July 2026) lets you require basic fundamental quality bars on the same screen that carries the RSI tile — use it to strip structurally impaired names out of the bullish pool before acting on any oversold fire.
  • Let the table set the exit discipline — Compare the 20d and 60d columns in the current tables. If a side's alpha grows with horizon, the effect is a slow drift rather than a sharp crack, and a time-stop that holds the full window captures more of it than a profit target that exits on the first bounce. If the columns are flat or reverse, shorten holds accordingly. Re-check after each backtest refresh — horizon behavior is exactly the kind of parameter that rots.

See also Why technical-only signals don't survive on their own for the broader argument.

5 · Before you act — a 5-point checklist

  1. Normal trading day? Rule out earnings (within ±3 days), ex-dividend, or known corporate-action dates — the signal is almost certainly reading noise, not momentum, in those windows.
  2. Where is price vs its own 50 / 200 DMA? A mean-reversion signal firing against the long-term trend (e.g. oversold in a clean uptrend) is much more reliable than one firing with it.
  3. What's the sector breadth doing? An isolated signal in a broadly down-trending sector is a lower-confidence setup than one firing with the rest of its peer group.
  4. Is ADV20 enough for your size? If the trigger is on a $500M name and you want to move $1M notional, you're the tape. Consider adv20d ≥ 5% of your intended position.
  5. What invalidates you? Define a price level (for longs: a close below the trigger-day low; for shorts: close above the trigger-day high) and honor it. The backtest alpha is an average; any one trade can be at either tail.

Execution notes

Entry convention in the tables above is next-day open (open T+1 after the trigger prints on the close). Read direction from the live data, not from folklore: under the single-sided long convention used here, a bearish signal 'works' when its post-trigger alpha is negative — the stock lags the benchmark after firing. If a side clears the random-date permutation null in the current tables, it can serve as a screen tile in that direction; if it does not, treat its fires as context rather than a trade prompt. Choose the holding horizon from the at-a-glance columns the same way rather than assuming a default. All of this describes historical tendency, not a recommendation.

See a live screen of this signal

Curated Daily Reports that screen for this signal, refreshed at the open, midday and close — free to view, no account needed: